That was it! The odd thing is all the cameras are the same model, and the issue occurred on my first 2 cameras on both instances I run of frigate the rest of the cameras were fine. Once i added the FPS setting, the massages "1600711741.614265 is not in memory store” also stopped.
Try increasing the max area settings for the camera. This is what I have mine set to max_area: 1000000. Without increasing the max area, I would miss quite a bit.
Crop and resize the best.jpg endpoint
Thank you, this is very helpfull!
thanks mate, i have set that just now, maybe today i wont have traffic on that backroad so probably i see and test more tomorrow.
except the shipping is as much as the product which seems a bit excessive
Latest frigate version from docker, even the object detection doesnt work properly. By the way the max_area: 1000000 did not help at all, still missing cars, trucks. I am waiting Blake’s answer.
Thanks for the great work ! The detection is working for people and dogs, but I have something similar as @rancho, cars or trucks are not detected. CPU usage is also very high when viewing stats in Portainer (usually ~140%). I’m not using zones or clips and I have frigate docker image v0.6.0. Camera feed is a 360°.
@WhistleMaster can you post your config? I forgot to mention in the changelog that cars/trucks are no longer tracked by default.
In my use case (just wanting to detect people) I can confirm the great capacity to skip many of the past false positives of the 0.6.0 release. The only two false positives I got was about a cat detected by Coral as person with a very high score (more than 0.90). As long as I understand the new Frigate heuristic confirm the detection only if the object is trucked (with high scores) in multiple frames: this saved me from “moving green” and “flight insects”. The “pussy problem” could be related to limitation of the Coral model. Waiting for a better model (https://github.com/google/automl/issues/11).
1 Like
Sure !
web_port: 5000
tensorflow_device: usb
mqtt:
host: <MQTT_IP>
topic_prefix: frigate
user: <MQTT_USER>
password: <MQTT_PWD>
save_clips:
max_seconds: 300
objects:
track:
- person
- car
- truck
- cat
- dog
filters:
person:
min_area: 5000
max_area: 100000
min_score: 0.5
threshold: 0.85
cameras:
front:
ffmpeg:
input: rtsp://<CAM_USER>:<CAM_PWD>@<CAM_IP>/<CAM_PATH>
fps: 10
take_frame: 1
best_image_timeout: 60
save_clips:
enabled: False
pre_capture: 30
snapshots:
show_timestamp: True
draw_zones: False
Yes. I am going to look at custom models in v0.9.0
1 Like
Hey guys, can anyone help me get going? I am very new to docker (I just downloaded it for this project). I am running ubuntu on a laptop. I have docker installed and I pulled the docker container to the computer. I dont know what to do next. I tried to run the container with the commands from the Github “getting started” section but I get errors. Does anyone have a guide for a complete newbie to this stuff? I have the USB Coral ready and everything just need to get it started.
Thanks for any help.
1 Like
your goal is the same as ours. To improve it.
It would be great to set the first milestone to improve the object detection and second milestone for the NVR function.
Funny thing both are down to the same thing. The hardware. (besides firgate)
In order to improve the detection you probably need different models and probably Tensorflow instead of TF light.
For object detection hardware acceleration you have Coral which is good here you need also to offload the CPU from decoding the stream, so you need hardware acceleration too.
Then what hardware? Really not much option left at all. Probably Jetson Nano is the only option that could do GPU accelerated object detection and video stream acceleration too. Besides Jetson Nano, of course next step is a real GPU like 1080Ti or 2070. Such build would cover the NVR function too probably.
Maybe a Jetson and Coral USB combo could could be enough for a few cam setup, this would worth to test if supported by frigate. Otherwise for really good accuracy we all end up with a full Corei7 PC with NVIDA GPU which of course use lot more electricty than a small Jetson - Coral build.
So thats the current technology limits and different routes to go ahead. I have been testing similar projects, same story there too. Probably this would be the best hardware for all this.: https://www.youtube.com/watch?v=lcT5jLl538Q
-
you need to have a config.yml file configured and saved in a config folder. lets just assume you saved it in /opt/frigate/config in your ubuntu laptop. Refer below link if you haven’t configured your config.yml file.
https://github.com/blakeblackshear/frigate/blob/309c0dcda3279ea06645aa6590ca055c615d8912/config/config.example.yml -
try to run the following docker run command to see if it works for you.
sudo docker run --name frigate --privileged --shm-size=512m -v /opt/frigate/config:/config:ro -v /etc/localtime:/etc/localtime:ro -v /dev/bus/usb:/dev/bus/usb -d -p 5000:5000 -e FRIGATE_RTSP_PASSWORD='password' blakeblackshear/frigate:stable
- run the following command to see the logs output of frigate
sudo docker logs frigate
1 Like
Are you using a Coral TPU?
So ever since my wife has decorated the front of our house with some fall colors I have been getting some of the plants triggering frigate. Thoughts on how I can correct his (besides moving the plants).
(upload://aO9tUJdnwcOyyWFqXagNHmZCpGM.jpeg) 
![Image004|640x480]
Yes I am, USB Coral.
Hi Blake, ever considered a 4D CNN running on batches of frames rather than object detection, motion and tracking. Just read an interesting paper:
ReMotENet: Efficient Relevant Motion Event Detection for Large-scale Home Surveillance Videos
Ruichi Yu, Hongcheng Wang, Larry S. Davis
Sadly though no code posted or parameters!
Increase your threshold, set a high min_size for person, or use a mask.
1 Like